Graph-Based Knowledge Representation and Semantic Alignment for Large Language Model-Driven Entity Recognition

Abstract

This study proposes a large language model-based entity recognition method enhanced with external knowledge graphs to address the limitations of traditional approaches in complex semantic environments, ambiguous entities, and knowledge-sparse scenarios. The method builds a multilayer framework consisting of text encoding, knowledge retrieval, graph-based representation, and semantic fusion. The large language model extracts deep semantic features, while the knowledge graph supplements external knowledge through structured nodes and relations, which strengthen reasoning in boundary detection and semantic disambiguation. A dynamic knowledge selection mechanism and a graph encoding module are designed to achieve deep alignment between textual information and knowledge representations, allowing the model to use graph substructures that are most relevant to the context during inference. Experiments on an open-domain entity recognition dataset, along with sensitivity analyses, demonstrate that the method achieves significant improvements in Accuracy, Precision, Recall, and F1 score. It is especially stable in rare entity cases, ambiguous expressions, and long-distance dependency settings. Further analyses of training size, class imbalance, distribution perturbation, and knowledge injection strength confirm the robustness and scalability of the framework under different data conditions and training environments. Overall, the method achieves deep integration of large language models and knowledge graphs in entity recognition and provides an effective path toward building more accurate, interpretable, and structure-aware recognition systems.